Reward-hacked GPT-4.1 models evaded shutdown in simulated evaluations
Review RAT-163 · Catalogue RAT-163
School of Reward Hacks trains models on intentionally reward-hacking examples then tests artificial shutdown responses. No actual persistence or organizational sabotage outside the evaluation is established.
Reviewed: · Rule: roguewatch-organizational-harm-v1
Dates and discovery
- Incident date
- Not disclosed
- Added to research / record
- Posted to RogueWatch
- No linked public incident
- First observed
- 2026-10-30T13:13:41.706307Z UTC
- Discovery channel
- User-supplied comparison audit
- Discovery source
- https://rogueaitracker.com/incidents/reward-hacked-gpt-4-1-models-evaded-shutdown/
Dates use America/Los_Angeles. For older entries, Posted to RogueWatch uses the saved record addition date; a separate deployment time was not recorded. Future posting timestamps record first public qualification. Source publication and incident occurrence are separate.
Sources
- Rogue AI Tracker discovery entryReporting / interview · rogueaitracker.com
- arxiv.orgOriginal account · arxiv.org
Decision history
- · Case-by-case comparison audit
RAT-163: out-of-scope. School of Reward Hacks trains models on intentionally reward-hacking examples then tests artificial shutdown responses. No actual persistence or organizational sabotage outside the evaluation is established. Source checks: read: https://arxiv.org/abs/2508.17511. School of Reward Hacks trains models on intentionally reward-hacking examples then tests artificial shutdown responses. No actual persistence or organizational sabotage outside the evaluation is established.
Evidence source
A review decision applies only to the named scope. Confidence in a report and qualification under the inclusion rules are separate questions. Review records do not add incidents to the globe or leaderboard.